xref: /freebsd/sys/kern/kern_cpuset.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright (c) 2008,  Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Copyright (c) 2008 Nokia Corporation
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice unmodified, this list of conditions, and the following
13  *    disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_ddb.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/sysproto.h>
39 #include <sys/jail.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
44 #include <sys/priv.h>
45 #include <sys/proc.h>
46 #include <sys/refcount.h>
47 #include <sys/sched.h>
48 #include <sys/smp.h>
49 #include <sys/syscallsubr.h>
50 #include <sys/capsicum.h>
51 #include <sys/cpuset.h>
52 #include <sys/sx.h>
53 #include <sys/queue.h>
54 #include <sys/libkern.h>
55 #include <sys/limits.h>
56 #include <sys/bus.h>
57 #include <sys/interrupt.h>
58 
59 #include <vm/uma.h>
60 #include <vm/vm.h>
61 #include <vm/vm_page.h>
62 #include <vm/vm_param.h>
63 
64 #ifdef DDB
65 #include <ddb/ddb.h>
66 #endif /* DDB */
67 
68 /*
69  * cpusets provide a mechanism for creating and manipulating sets of
70  * processors for the purpose of constraining the scheduling of threads to
71  * specific processors.
72  *
73  * Each process belongs to an identified set, by default this is set 1.  Each
74  * thread may further restrict the cpus it may run on to a subset of this
75  * named set.  This creates an anonymous set which other threads and processes
76  * may not join by number.
77  *
78  * The named set is referred to herein as the 'base' set to avoid ambiguity.
79  * This set is usually a child of a 'root' set while the anonymous set may
80  * simply be referred to as a mask.  In the syscall api these are referred to
81  * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
82  *
83  * Threads inherit their set from their creator whether it be anonymous or
84  * not.  This means that anonymous sets are immutable because they may be
85  * shared.  To modify an anonymous set a new set is created with the desired
86  * mask and the same parent as the existing anonymous set.  This gives the
87  * illusion of each thread having a private mask.
88  *
89  * Via the syscall apis a user may ask to retrieve or modify the root, base,
90  * or mask that is discovered via a pid, tid, or setid.  Modifying a set
91  * modifies all numbered and anonymous child sets to comply with the new mask.
92  * Modifying a pid or tid's mask applies only to that tid but must still
93  * exist within the assigned parent set.
94  *
95  * A thread may not be assigned to a group separate from other threads in
96  * the process.  This is to remove ambiguity when the setid is queried with
97  * a pid argument.  There is no other technical limitation.
98  *
99  * This somewhat complex arrangement is intended to make it easy for
100  * applications to query available processors and bind their threads to
101  * specific processors while also allowing administrators to dynamically
102  * reprovision by changing sets which apply to groups of processes.
103  *
104  * A simple application should not concern itself with sets at all and
105  * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
106  * meaning 'curthread'.  It may query available cpus for that tid with a
107  * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
108  */
109 static uma_zone_t cpuset_zone;
110 static struct mtx cpuset_lock;
111 static struct setlist cpuset_ids;
112 static struct unrhdr *cpuset_unr;
113 static struct cpuset *cpuset_zero, *cpuset_default;
114 
115 /* Return the size of cpuset_t at the kernel level */
116 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
117     SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
118 
119 cpuset_t *cpuset_root;
120 cpuset_t cpuset_domain[MAXMEMDOM];
121 
122 /*
123  * Acquire a reference to a cpuset, all pointers must be tracked with refs.
124  */
125 struct cpuset *
126 cpuset_ref(struct cpuset *set)
127 {
128 
129 	refcount_acquire(&set->cs_ref);
130 	return (set);
131 }
132 
133 /*
134  * Walks up the tree from 'set' to find the root.  Returns the root
135  * referenced.
136  */
137 static struct cpuset *
138 cpuset_refroot(struct cpuset *set)
139 {
140 
141 	for (; set->cs_parent != NULL; set = set->cs_parent)
142 		if (set->cs_flags & CPU_SET_ROOT)
143 			break;
144 	cpuset_ref(set);
145 
146 	return (set);
147 }
148 
149 /*
150  * Find the first non-anonymous set starting from 'set'.  Returns this set
151  * referenced.  May return the passed in set with an extra ref if it is
152  * not anonymous.
153  */
154 static struct cpuset *
155 cpuset_refbase(struct cpuset *set)
156 {
157 
158 	if (set->cs_id == CPUSET_INVALID)
159 		set = set->cs_parent;
160 	cpuset_ref(set);
161 
162 	return (set);
163 }
164 
165 /*
166  * Release a reference in a context where it is safe to allocate.
167  */
168 void
169 cpuset_rel(struct cpuset *set)
170 {
171 	cpusetid_t id;
172 
173 	if (refcount_release(&set->cs_ref) == 0)
174 		return;
175 	mtx_lock_spin(&cpuset_lock);
176 	LIST_REMOVE(set, cs_siblings);
177 	id = set->cs_id;
178 	if (id != CPUSET_INVALID)
179 		LIST_REMOVE(set, cs_link);
180 	mtx_unlock_spin(&cpuset_lock);
181 	cpuset_rel(set->cs_parent);
182 	uma_zfree(cpuset_zone, set);
183 	if (id != CPUSET_INVALID)
184 		free_unr(cpuset_unr, id);
185 }
186 
187 /*
188  * Deferred release must be used when in a context that is not safe to
189  * allocate/free.  This places any unreferenced sets on the list 'head'.
190  */
191 static void
192 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
193 {
194 
195 	if (refcount_release(&set->cs_ref) == 0)
196 		return;
197 	mtx_lock_spin(&cpuset_lock);
198 	LIST_REMOVE(set, cs_siblings);
199 	if (set->cs_id != CPUSET_INVALID)
200 		LIST_REMOVE(set, cs_link);
201 	LIST_INSERT_HEAD(head, set, cs_link);
202 	mtx_unlock_spin(&cpuset_lock);
203 }
204 
205 /*
206  * Complete a deferred release.  Removes the set from the list provided to
207  * cpuset_rel_defer.
208  */
209 static void
210 cpuset_rel_complete(struct cpuset *set)
211 {
212 	LIST_REMOVE(set, cs_link);
213 	cpuset_rel(set->cs_parent);
214 	uma_zfree(cpuset_zone, set);
215 }
216 
217 /*
218  * Find a set based on an id.  Returns it with a ref.
219  */
220 static struct cpuset *
221 cpuset_lookup(cpusetid_t setid, struct thread *td)
222 {
223 	struct cpuset *set;
224 
225 	if (setid == CPUSET_INVALID)
226 		return (NULL);
227 	mtx_lock_spin(&cpuset_lock);
228 	LIST_FOREACH(set, &cpuset_ids, cs_link)
229 		if (set->cs_id == setid)
230 			break;
231 	if (set)
232 		cpuset_ref(set);
233 	mtx_unlock_spin(&cpuset_lock);
234 
235 	KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
236 	if (set != NULL && jailed(td->td_ucred)) {
237 		struct cpuset *jset, *tset;
238 
239 		jset = td->td_ucred->cr_prison->pr_cpuset;
240 		for (tset = set; tset != NULL; tset = tset->cs_parent)
241 			if (tset == jset)
242 				break;
243 		if (tset == NULL) {
244 			cpuset_rel(set);
245 			set = NULL;
246 		}
247 	}
248 
249 	return (set);
250 }
251 
252 /*
253  * Create a set in the space provided in 'set' with the provided parameters.
254  * The set is returned with a single ref.  May return EDEADLK if the set
255  * will have no valid cpu based on restrictions from the parent.
256  */
257 static int
258 _cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask,
259     cpusetid_t id)
260 {
261 
262 	if (!CPU_OVERLAP(&parent->cs_mask, mask))
263 		return (EDEADLK);
264 	CPU_COPY(mask, &set->cs_mask);
265 	LIST_INIT(&set->cs_children);
266 	refcount_init(&set->cs_ref, 1);
267 	set->cs_flags = 0;
268 	mtx_lock_spin(&cpuset_lock);
269 	CPU_AND(&set->cs_mask, &parent->cs_mask);
270 	set->cs_id = id;
271 	set->cs_parent = cpuset_ref(parent);
272 	LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
273 	if (set->cs_id != CPUSET_INVALID)
274 		LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
275 	mtx_unlock_spin(&cpuset_lock);
276 
277 	return (0);
278 }
279 
280 /*
281  * Create a new non-anonymous set with the requested parent and mask.  May
282  * return failures if the mask is invalid or a new number can not be
283  * allocated.
284  */
285 static int
286 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
287 {
288 	struct cpuset *set;
289 	cpusetid_t id;
290 	int error;
291 
292 	id = alloc_unr(cpuset_unr);
293 	if (id == -1)
294 		return (ENFILE);
295 	*setp = set = uma_zalloc(cpuset_zone, M_WAITOK);
296 	error = _cpuset_create(set, parent, mask, id);
297 	if (error == 0)
298 		return (0);
299 	free_unr(cpuset_unr, id);
300 	uma_zfree(cpuset_zone, set);
301 
302 	return (error);
303 }
304 
305 /*
306  * Recursively check for errors that would occur from applying mask to
307  * the tree of sets starting at 'set'.  Checks for sets that would become
308  * empty as well as RDONLY flags.
309  */
310 static int
311 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask)
312 {
313 	struct cpuset *nset;
314 	cpuset_t newmask;
315 	int error;
316 
317 	mtx_assert(&cpuset_lock, MA_OWNED);
318 	if (set->cs_flags & CPU_SET_RDONLY)
319 		return (EPERM);
320 	if (check_mask) {
321 		if (!CPU_OVERLAP(&set->cs_mask, mask))
322 			return (EDEADLK);
323 		CPU_COPY(&set->cs_mask, &newmask);
324 		CPU_AND(&newmask, mask);
325 	} else
326 		CPU_COPY(mask, &newmask);
327 	error = 0;
328 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
329 		if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
330 			break;
331 	return (error);
332 }
333 
334 /*
335  * Applies the mask 'mask' without checking for empty sets or permissions.
336  */
337 static void
338 cpuset_update(struct cpuset *set, cpuset_t *mask)
339 {
340 	struct cpuset *nset;
341 
342 	mtx_assert(&cpuset_lock, MA_OWNED);
343 	CPU_AND(&set->cs_mask, mask);
344 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
345 		cpuset_update(nset, &set->cs_mask);
346 
347 	return;
348 }
349 
350 /*
351  * Modify the set 'set' to use a copy of the mask provided.  Apply this new
352  * mask to restrict all children in the tree.  Checks for validity before
353  * applying the changes.
354  */
355 static int
356 cpuset_modify(struct cpuset *set, cpuset_t *mask)
357 {
358 	struct cpuset *root;
359 	int error;
360 
361 	error = priv_check(curthread, PRIV_SCHED_CPUSET);
362 	if (error)
363 		return (error);
364 	/*
365 	 * In case we are called from within the jail
366 	 * we do not allow modifying the dedicated root
367 	 * cpuset of the jail but may still allow to
368 	 * change child sets.
369 	 */
370 	if (jailed(curthread->td_ucred) &&
371 	    set->cs_flags & CPU_SET_ROOT)
372 		return (EPERM);
373 	/*
374 	 * Verify that we have access to this set of
375 	 * cpus.
376 	 */
377 	root = set->cs_parent;
378 	if (root && !CPU_SUBSET(&root->cs_mask, mask))
379 		return (EINVAL);
380 	mtx_lock_spin(&cpuset_lock);
381 	error = cpuset_testupdate(set, mask, 0);
382 	if (error)
383 		goto out;
384 	CPU_COPY(mask, &set->cs_mask);
385 	cpuset_update(set, mask);
386 out:
387 	mtx_unlock_spin(&cpuset_lock);
388 
389 	return (error);
390 }
391 
392 /*
393  * Resolve the 'which' parameter of several cpuset apis.
394  *
395  * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid.  Also
396  * checks for permission via p_cansched().
397  *
398  * For WHICH_SET returns a valid set with a new reference.
399  *
400  * -1 may be supplied for any argument to mean the current proc/thread or
401  * the base set of the current thread.  May fail with ESRCH/EPERM.
402  */
403 int
404 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
405     struct cpuset **setp)
406 {
407 	struct cpuset *set;
408 	struct thread *td;
409 	struct proc *p;
410 	int error;
411 
412 	*pp = p = NULL;
413 	*tdp = td = NULL;
414 	*setp = set = NULL;
415 	switch (which) {
416 	case CPU_WHICH_PID:
417 		if (id == -1) {
418 			PROC_LOCK(curproc);
419 			p = curproc;
420 			break;
421 		}
422 		if ((p = pfind(id)) == NULL)
423 			return (ESRCH);
424 		break;
425 	case CPU_WHICH_TID:
426 		if (id == -1) {
427 			PROC_LOCK(curproc);
428 			p = curproc;
429 			td = curthread;
430 			break;
431 		}
432 		td = tdfind(id, -1);
433 		if (td == NULL)
434 			return (ESRCH);
435 		p = td->td_proc;
436 		break;
437 	case CPU_WHICH_CPUSET:
438 		if (id == -1) {
439 			thread_lock(curthread);
440 			set = cpuset_refbase(curthread->td_cpuset);
441 			thread_unlock(curthread);
442 		} else
443 			set = cpuset_lookup(id, curthread);
444 		if (set) {
445 			*setp = set;
446 			return (0);
447 		}
448 		return (ESRCH);
449 	case CPU_WHICH_JAIL:
450 	{
451 		/* Find `set' for prison with given id. */
452 		struct prison *pr;
453 
454 		sx_slock(&allprison_lock);
455 		pr = prison_find_child(curthread->td_ucred->cr_prison, id);
456 		sx_sunlock(&allprison_lock);
457 		if (pr == NULL)
458 			return (ESRCH);
459 		cpuset_ref(pr->pr_cpuset);
460 		*setp = pr->pr_cpuset;
461 		mtx_unlock(&pr->pr_mtx);
462 		return (0);
463 	}
464 	case CPU_WHICH_IRQ:
465 	case CPU_WHICH_DOMAIN:
466 		return (0);
467 	default:
468 		return (EINVAL);
469 	}
470 	error = p_cansched(curthread, p);
471 	if (error) {
472 		PROC_UNLOCK(p);
473 		return (error);
474 	}
475 	if (td == NULL)
476 		td = FIRST_THREAD_IN_PROC(p);
477 	*pp = p;
478 	*tdp = td;
479 	return (0);
480 }
481 
482 /*
483  * Create an anonymous set with the provided mask in the space provided by
484  * 'fset'.  If the passed in set is anonymous we use its parent otherwise
485  * the new set is a child of 'set'.
486  */
487 static int
488 cpuset_shadow(struct cpuset *set, struct cpuset *fset, const cpuset_t *mask)
489 {
490 	struct cpuset *parent;
491 
492 	if (set->cs_id == CPUSET_INVALID)
493 		parent = set->cs_parent;
494 	else
495 		parent = set;
496 	if (!CPU_SUBSET(&parent->cs_mask, mask))
497 		return (EDEADLK);
498 	return (_cpuset_create(fset, parent, mask, CPUSET_INVALID));
499 }
500 
501 /*
502  * Handle two cases for replacing the base set or mask of an entire process.
503  *
504  * 1) Set is non-null and mask is null.  This reparents all anonymous sets
505  *    to the provided set and replaces all non-anonymous td_cpusets with the
506  *    provided set.
507  * 2) Mask is non-null and set is null.  This replaces or creates anonymous
508  *    sets for every thread with the existing base as a parent.
509  *
510  * This is overly complicated because we can't allocate while holding a
511  * spinlock and spinlocks must be held while changing and examining thread
512  * state.
513  */
514 static int
515 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask)
516 {
517 	struct setlist freelist;
518 	struct setlist droplist;
519 	struct cpuset *tdset;
520 	struct cpuset *nset;
521 	struct thread *td;
522 	struct proc *p;
523 	int threads;
524 	int nfree;
525 	int error;
526 
527 	/*
528 	 * The algorithm requires two passes due to locking considerations.
529 	 *
530 	 * 1) Lookup the process and acquire the locks in the required order.
531 	 * 2) If enough cpusets have not been allocated release the locks and
532 	 *    allocate them.  Loop.
533 	 */
534 	LIST_INIT(&freelist);
535 	LIST_INIT(&droplist);
536 	nfree = 0;
537 	for (;;) {
538 		error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
539 		if (error)
540 			goto out;
541 		if (nfree >= p->p_numthreads)
542 			break;
543 		threads = p->p_numthreads;
544 		PROC_UNLOCK(p);
545 		for (; nfree < threads; nfree++) {
546 			nset = uma_zalloc(cpuset_zone, M_WAITOK);
547 			LIST_INSERT_HEAD(&freelist, nset, cs_link);
548 		}
549 	}
550 	PROC_LOCK_ASSERT(p, MA_OWNED);
551 	/*
552 	 * Now that the appropriate locks are held and we have enough cpusets,
553 	 * make sure the operation will succeed before applying changes.  The
554 	 * proc lock prevents td_cpuset from changing between calls.
555 	 */
556 	error = 0;
557 	FOREACH_THREAD_IN_PROC(p, td) {
558 		thread_lock(td);
559 		tdset = td->td_cpuset;
560 		/*
561 		 * Verify that a new mask doesn't specify cpus outside of
562 		 * the set the thread is a member of.
563 		 */
564 		if (mask) {
565 			if (tdset->cs_id == CPUSET_INVALID)
566 				tdset = tdset->cs_parent;
567 			if (!CPU_SUBSET(&tdset->cs_mask, mask))
568 				error = EDEADLK;
569 		/*
570 		 * Verify that a new set won't leave an existing thread
571 		 * mask without a cpu to run on.  It can, however, restrict
572 		 * the set.
573 		 */
574 		} else if (tdset->cs_id == CPUSET_INVALID) {
575 			if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask))
576 				error = EDEADLK;
577 		}
578 		thread_unlock(td);
579 		if (error)
580 			goto unlock_out;
581 	}
582 	/*
583 	 * Replace each thread's cpuset while using deferred release.  We
584 	 * must do this because the thread lock must be held while operating
585 	 * on the thread and this limits the type of operations allowed.
586 	 */
587 	FOREACH_THREAD_IN_PROC(p, td) {
588 		thread_lock(td);
589 		/*
590 		 * If we presently have an anonymous set or are applying a
591 		 * mask we must create an anonymous shadow set.  That is
592 		 * either parented to our existing base or the supplied set.
593 		 *
594 		 * If we have a base set with no anonymous shadow we simply
595 		 * replace it outright.
596 		 */
597 		tdset = td->td_cpuset;
598 		if (tdset->cs_id == CPUSET_INVALID || mask) {
599 			nset = LIST_FIRST(&freelist);
600 			LIST_REMOVE(nset, cs_link);
601 			if (mask)
602 				error = cpuset_shadow(tdset, nset, mask);
603 			else
604 				error = _cpuset_create(nset, set,
605 				    &tdset->cs_mask, CPUSET_INVALID);
606 			if (error) {
607 				LIST_INSERT_HEAD(&freelist, nset, cs_link);
608 				thread_unlock(td);
609 				break;
610 			}
611 		} else
612 			nset = cpuset_ref(set);
613 		cpuset_rel_defer(&droplist, tdset);
614 		td->td_cpuset = nset;
615 		sched_affinity(td);
616 		thread_unlock(td);
617 	}
618 unlock_out:
619 	PROC_UNLOCK(p);
620 out:
621 	while ((nset = LIST_FIRST(&droplist)) != NULL)
622 		cpuset_rel_complete(nset);
623 	while ((nset = LIST_FIRST(&freelist)) != NULL) {
624 		LIST_REMOVE(nset, cs_link);
625 		uma_zfree(cpuset_zone, nset);
626 	}
627 	return (error);
628 }
629 
630 /*
631  * Return a string representing a valid layout for a cpuset_t object.
632  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
633  */
634 char *
635 cpusetobj_strprint(char *buf, const cpuset_t *set)
636 {
637 	char *tbuf;
638 	size_t i, bytesp, bufsiz;
639 
640 	tbuf = buf;
641 	bytesp = 0;
642 	bufsiz = CPUSETBUFSIZ;
643 
644 	for (i = 0; i < (_NCPUWORDS - 1); i++) {
645 		bytesp = snprintf(tbuf, bufsiz, "%lx,", set->__bits[i]);
646 		bufsiz -= bytesp;
647 		tbuf += bytesp;
648 	}
649 	snprintf(tbuf, bufsiz, "%lx", set->__bits[_NCPUWORDS - 1]);
650 	return (buf);
651 }
652 
653 /*
654  * Build a valid cpuset_t object from a string representation.
655  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
656  */
657 int
658 cpusetobj_strscan(cpuset_t *set, const char *buf)
659 {
660 	u_int nwords;
661 	int i, ret;
662 
663 	if (strlen(buf) > CPUSETBUFSIZ - 1)
664 		return (-1);
665 
666 	/* Allow to pass a shorter version of the mask when necessary. */
667 	nwords = 1;
668 	for (i = 0; buf[i] != '\0'; i++)
669 		if (buf[i] == ',')
670 			nwords++;
671 	if (nwords > _NCPUWORDS)
672 		return (-1);
673 
674 	CPU_ZERO(set);
675 	for (i = 0; i < (nwords - 1); i++) {
676 		ret = sscanf(buf, "%lx,", &set->__bits[i]);
677 		if (ret == 0 || ret == -1)
678 			return (-1);
679 		buf = strstr(buf, ",");
680 		if (buf == NULL)
681 			return (-1);
682 		buf++;
683 	}
684 	ret = sscanf(buf, "%lx", &set->__bits[nwords - 1]);
685 	if (ret == 0 || ret == -1)
686 		return (-1);
687 	return (0);
688 }
689 
690 /*
691  * Apply an anonymous mask to a single thread.
692  */
693 int
694 cpuset_setthread(lwpid_t id, cpuset_t *mask)
695 {
696 	struct cpuset *nset;
697 	struct cpuset *set;
698 	struct thread *td;
699 	struct proc *p;
700 	int error;
701 
702 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
703 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
704 	if (error)
705 		goto out;
706 	set = NULL;
707 	thread_lock(td);
708 	error = cpuset_shadow(td->td_cpuset, nset, mask);
709 	if (error == 0) {
710 		set = td->td_cpuset;
711 		td->td_cpuset = nset;
712 		sched_affinity(td);
713 		nset = NULL;
714 	}
715 	thread_unlock(td);
716 	PROC_UNLOCK(p);
717 	if (set)
718 		cpuset_rel(set);
719 out:
720 	if (nset)
721 		uma_zfree(cpuset_zone, nset);
722 	return (error);
723 }
724 
725 /*
726  * Apply new cpumask to the ithread.
727  */
728 int
729 cpuset_setithread(lwpid_t id, int cpu)
730 {
731 	struct cpuset *nset, *rset;
732 	struct cpuset *parent, *old_set;
733 	struct thread *td;
734 	struct proc *p;
735 	cpusetid_t cs_id;
736 	cpuset_t mask;
737 	int error;
738 
739 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
740 	rset = uma_zalloc(cpuset_zone, M_WAITOK);
741 	cs_id = CPUSET_INVALID;
742 
743 	CPU_ZERO(&mask);
744 	if (cpu == NOCPU)
745 		CPU_COPY(cpuset_root, &mask);
746 	else
747 		CPU_SET(cpu, &mask);
748 
749 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &old_set);
750 	if (error != 0 || ((cs_id = alloc_unr(cpuset_unr)) == CPUSET_INVALID))
751 		goto out;
752 
753 	/* cpuset_which() returns with PROC_LOCK held. */
754 	old_set = td->td_cpuset;
755 
756 	if (cpu == NOCPU) {
757 
758 		/*
759 		 * roll back to default set. We're not using cpuset_shadow()
760 		 * here because we can fail CPU_SUBSET() check. This can happen
761 		 * if default set does not contain all CPUs.
762 		 */
763 		error = _cpuset_create(nset, cpuset_default, &mask,
764 		    CPUSET_INVALID);
765 
766 		goto applyset;
767 	}
768 
769 	if (old_set->cs_id == 1 || (old_set->cs_id == CPUSET_INVALID &&
770 	    old_set->cs_parent->cs_id == 1)) {
771 
772 		/*
773 		 * Current set is either default (1) or
774 		 * shadowed version of default set.
775 		 *
776 		 * Allocate new root set to be able to shadow it
777 		 * with any mask.
778 		 */
779 		error = _cpuset_create(rset, cpuset_zero,
780 		    &cpuset_zero->cs_mask, cs_id);
781 		if (error != 0) {
782 			PROC_UNLOCK(p);
783 			goto out;
784 		}
785 		rset->cs_flags |= CPU_SET_ROOT;
786 		parent = rset;
787 		rset = NULL;
788 		cs_id = CPUSET_INVALID;
789 	} else {
790 		/* Assume existing set was already allocated by previous call */
791 		parent = old_set;
792 		old_set = NULL;
793 	}
794 
795 	error = cpuset_shadow(parent, nset, &mask);
796 applyset:
797 	if (error == 0) {
798 		thread_lock(td);
799 		td->td_cpuset = nset;
800 		sched_affinity(td);
801 		thread_unlock(td);
802 		nset = NULL;
803 	} else
804 		old_set = NULL;
805 	PROC_UNLOCK(p);
806 	if (old_set != NULL)
807 		cpuset_rel(old_set);
808 out:
809 	if (nset != NULL)
810 		uma_zfree(cpuset_zone, nset);
811 	if (rset != NULL)
812 		uma_zfree(cpuset_zone, rset);
813 	if (cs_id != CPUSET_INVALID)
814 		free_unr(cpuset_unr, cs_id);
815 	return (error);
816 }
817 
818 
819 /*
820  * Creates system-wide cpusets and the cpuset for thread0 including two
821  * sets:
822  *
823  * 0 - The root set which should represent all valid processors in the
824  *     system.  It is initially created with a mask of all processors
825  *     because we don't know what processors are valid until cpuset_init()
826  *     runs.  This set is immutable.
827  * 1 - The default set which all processes are a member of until changed.
828  *     This allows an administrator to move all threads off of given cpus to
829  *     dedicate them to high priority tasks or save power etc.
830  */
831 struct cpuset *
832 cpuset_thread0(void)
833 {
834 	struct cpuset *set;
835 	int error, i;
836 
837 	cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
838 	    NULL, NULL, UMA_ALIGN_PTR, 0);
839 	mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
840 
841 	/*
842 	 * Create the root system set for the whole machine.  Doesn't use
843 	 * cpuset_create() due to NULL parent.
844 	 */
845 	set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
846 	CPU_FILL(&set->cs_mask);
847 	LIST_INIT(&set->cs_children);
848 	LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
849 	set->cs_ref = 1;
850 	set->cs_flags = CPU_SET_ROOT;
851 	cpuset_zero = set;
852 	cpuset_root = &set->cs_mask;
853 
854 	/*
855 	 * Now derive a default, modifiable set from that to give out.
856 	 */
857 	set = uma_zalloc(cpuset_zone, M_WAITOK);
858 	error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1);
859 	KASSERT(error == 0, ("Error creating default set: %d\n", error));
860 	cpuset_default = set;
861 
862 	/*
863 	 * Initialize the unit allocator. 0 and 1 are allocated above.
864 	 */
865 	cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
866 
867 	/*
868 	 * If MD code has not initialized per-domain cpusets, place all
869 	 * CPUs in domain 0.
870 	 */
871 	for (i = 0; i < MAXMEMDOM; i++)
872 		if (!CPU_EMPTY(&cpuset_domain[i]))
873 			goto domains_set;
874 	CPU_COPY(&all_cpus, &cpuset_domain[0]);
875 domains_set:
876 
877 	return (set);
878 }
879 
880 /*
881  * Create a cpuset, which would be cpuset_create() but
882  * mark the new 'set' as root.
883  *
884  * We are not going to reparent the td to it.  Use cpuset_setproc_update_set()
885  * for that.
886  *
887  * In case of no error, returns the set in *setp locked with a reference.
888  */
889 int
890 cpuset_create_root(struct prison *pr, struct cpuset **setp)
891 {
892 	struct cpuset *set;
893 	int error;
894 
895 	KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
896 	KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
897 
898 	error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
899 	if (error)
900 		return (error);
901 
902 	KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data",
903 	    __func__, __LINE__));
904 
905 	/* Mark the set as root. */
906 	set = *setp;
907 	set->cs_flags |= CPU_SET_ROOT;
908 
909 	return (0);
910 }
911 
912 int
913 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
914 {
915 	int error;
916 
917 	KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
918 	KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
919 
920 	cpuset_ref(set);
921 	error = cpuset_setproc(p->p_pid, set, NULL);
922 	if (error)
923 		return (error);
924 	cpuset_rel(set);
925 	return (0);
926 }
927 
928 /*
929  * This is called once the final set of system cpus is known.  Modifies
930  * the root set and all children and mark the root read-only.
931  */
932 static void
933 cpuset_init(void *arg)
934 {
935 	cpuset_t mask;
936 
937 	mask = all_cpus;
938 	if (cpuset_modify(cpuset_zero, &mask))
939 		panic("Can't set initial cpuset mask.\n");
940 	cpuset_zero->cs_flags |= CPU_SET_RDONLY;
941 }
942 SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL);
943 
944 #ifndef _SYS_SYSPROTO_H_
945 struct cpuset_args {
946 	cpusetid_t	*setid;
947 };
948 #endif
949 int
950 sys_cpuset(struct thread *td, struct cpuset_args *uap)
951 {
952 	struct cpuset *root;
953 	struct cpuset *set;
954 	int error;
955 
956 	thread_lock(td);
957 	root = cpuset_refroot(td->td_cpuset);
958 	thread_unlock(td);
959 	error = cpuset_create(&set, root, &root->cs_mask);
960 	cpuset_rel(root);
961 	if (error)
962 		return (error);
963 	error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
964 	if (error == 0)
965 		error = cpuset_setproc(-1, set, NULL);
966 	cpuset_rel(set);
967 	return (error);
968 }
969 
970 #ifndef _SYS_SYSPROTO_H_
971 struct cpuset_setid_args {
972 	cpuwhich_t	which;
973 	id_t		id;
974 	cpusetid_t	setid;
975 };
976 #endif
977 int
978 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
979 {
980 
981 	return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
982 }
983 
984 int
985 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
986     id_t id, cpusetid_t setid)
987 {
988 	struct cpuset *set;
989 	int error;
990 
991 	/*
992 	 * Presently we only support per-process sets.
993 	 */
994 	if (which != CPU_WHICH_PID)
995 		return (EINVAL);
996 	set = cpuset_lookup(setid, td);
997 	if (set == NULL)
998 		return (ESRCH);
999 	error = cpuset_setproc(id, set, NULL);
1000 	cpuset_rel(set);
1001 	return (error);
1002 }
1003 
1004 #ifndef _SYS_SYSPROTO_H_
1005 struct cpuset_getid_args {
1006 	cpulevel_t	level;
1007 	cpuwhich_t	which;
1008 	id_t		id;
1009 	cpusetid_t	*setid;
1010 };
1011 #endif
1012 int
1013 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1014 {
1015 
1016 	return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1017 	    uap->setid));
1018 }
1019 
1020 int
1021 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1022     id_t id, cpusetid_t *setid)
1023 {
1024 	struct cpuset *nset;
1025 	struct cpuset *set;
1026 	struct thread *ttd;
1027 	struct proc *p;
1028 	cpusetid_t tmpid;
1029 	int error;
1030 
1031 	if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1032 		return (EINVAL);
1033 	error = cpuset_which(which, id, &p, &ttd, &set);
1034 	if (error)
1035 		return (error);
1036 	switch (which) {
1037 	case CPU_WHICH_TID:
1038 	case CPU_WHICH_PID:
1039 		thread_lock(ttd);
1040 		set = cpuset_refbase(ttd->td_cpuset);
1041 		thread_unlock(ttd);
1042 		PROC_UNLOCK(p);
1043 		break;
1044 	case CPU_WHICH_CPUSET:
1045 	case CPU_WHICH_JAIL:
1046 		break;
1047 	case CPU_WHICH_IRQ:
1048 	case CPU_WHICH_DOMAIN:
1049 		return (EINVAL);
1050 	}
1051 	switch (level) {
1052 	case CPU_LEVEL_ROOT:
1053 		nset = cpuset_refroot(set);
1054 		cpuset_rel(set);
1055 		set = nset;
1056 		break;
1057 	case CPU_LEVEL_CPUSET:
1058 		break;
1059 	case CPU_LEVEL_WHICH:
1060 		break;
1061 	}
1062 	tmpid = set->cs_id;
1063 	cpuset_rel(set);
1064 	if (error == 0)
1065 		error = copyout(&tmpid, setid, sizeof(id));
1066 
1067 	return (error);
1068 }
1069 
1070 #ifndef _SYS_SYSPROTO_H_
1071 struct cpuset_getaffinity_args {
1072 	cpulevel_t	level;
1073 	cpuwhich_t	which;
1074 	id_t		id;
1075 	size_t		cpusetsize;
1076 	cpuset_t	*mask;
1077 };
1078 #endif
1079 int
1080 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1081 {
1082 
1083 	return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1084 	    uap->id, uap->cpusetsize, uap->mask));
1085 }
1086 
1087 int
1088 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1089     id_t id, size_t cpusetsize, cpuset_t *maskp)
1090 {
1091 	struct thread *ttd;
1092 	struct cpuset *nset;
1093 	struct cpuset *set;
1094 	struct proc *p;
1095 	cpuset_t *mask;
1096 	int error;
1097 	size_t size;
1098 
1099 	if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1100 		return (ERANGE);
1101 	/* In Capability mode, you can only get your own CPU set. */
1102 	if (IN_CAPABILITY_MODE(td)) {
1103 	    if (level != CPU_LEVEL_WHICH)
1104 		return (ECAPMODE);
1105 	    if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1106 		return (ECAPMODE);
1107 	    if (id != -1)
1108 		return (ECAPMODE);
1109 	}
1110 	size = cpusetsize;
1111 	mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1112 	error = cpuset_which(which, id, &p, &ttd, &set);
1113 	if (error)
1114 		goto out;
1115 	switch (level) {
1116 	case CPU_LEVEL_ROOT:
1117 	case CPU_LEVEL_CPUSET:
1118 		switch (which) {
1119 		case CPU_WHICH_TID:
1120 		case CPU_WHICH_PID:
1121 			thread_lock(ttd);
1122 			set = cpuset_ref(ttd->td_cpuset);
1123 			thread_unlock(ttd);
1124 			break;
1125 		case CPU_WHICH_CPUSET:
1126 		case CPU_WHICH_JAIL:
1127 			break;
1128 		case CPU_WHICH_IRQ:
1129 		case CPU_WHICH_INTRHANDLER:
1130 		case CPU_WHICH_ITHREAD:
1131 		case CPU_WHICH_DOMAIN:
1132 			error = EINVAL;
1133 			goto out;
1134 		}
1135 		if (level == CPU_LEVEL_ROOT)
1136 			nset = cpuset_refroot(set);
1137 		else
1138 			nset = cpuset_refbase(set);
1139 		CPU_COPY(&nset->cs_mask, mask);
1140 		cpuset_rel(nset);
1141 		break;
1142 	case CPU_LEVEL_WHICH:
1143 		switch (which) {
1144 		case CPU_WHICH_TID:
1145 			thread_lock(ttd);
1146 			CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1147 			thread_unlock(ttd);
1148 			break;
1149 		case CPU_WHICH_PID:
1150 			FOREACH_THREAD_IN_PROC(p, ttd) {
1151 				thread_lock(ttd);
1152 				CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1153 				thread_unlock(ttd);
1154 			}
1155 			break;
1156 		case CPU_WHICH_CPUSET:
1157 		case CPU_WHICH_JAIL:
1158 			CPU_COPY(&set->cs_mask, mask);
1159 			break;
1160 		case CPU_WHICH_IRQ:
1161 		case CPU_WHICH_INTRHANDLER:
1162 		case CPU_WHICH_ITHREAD:
1163 			error = intr_getaffinity(id, which, mask);
1164 			break;
1165 		case CPU_WHICH_DOMAIN:
1166 			if (id < 0 || id >= MAXMEMDOM)
1167 				error = ESRCH;
1168 			else
1169 				CPU_COPY(&cpuset_domain[id], mask);
1170 			break;
1171 		}
1172 		break;
1173 	default:
1174 		error = EINVAL;
1175 		break;
1176 	}
1177 	if (set)
1178 		cpuset_rel(set);
1179 	if (p)
1180 		PROC_UNLOCK(p);
1181 	if (error == 0)
1182 		error = copyout(mask, maskp, size);
1183 out:
1184 	free(mask, M_TEMP);
1185 	return (error);
1186 }
1187 
1188 #ifndef _SYS_SYSPROTO_H_
1189 struct cpuset_setaffinity_args {
1190 	cpulevel_t	level;
1191 	cpuwhich_t	which;
1192 	id_t		id;
1193 	size_t		cpusetsize;
1194 	const cpuset_t	*mask;
1195 };
1196 #endif
1197 int
1198 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1199 {
1200 
1201 	return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1202 	    uap->id, uap->cpusetsize, uap->mask));
1203 }
1204 
1205 int
1206 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1207     id_t id, size_t cpusetsize, const cpuset_t *maskp)
1208 {
1209 	struct cpuset *nset;
1210 	struct cpuset *set;
1211 	struct thread *ttd;
1212 	struct proc *p;
1213 	cpuset_t *mask;
1214 	int error;
1215 
1216 	if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1217 		return (ERANGE);
1218 	/* In Capability mode, you can only set your own CPU set. */
1219 	if (IN_CAPABILITY_MODE(td)) {
1220 	    if (level != CPU_LEVEL_WHICH)
1221 		return (ECAPMODE);
1222 	    if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1223 		return (ECAPMODE);
1224 	    if (id != -1)
1225 		return (ECAPMODE);
1226 	}
1227 	mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1228 	error = copyin(maskp, mask, cpusetsize);
1229 	if (error)
1230 		goto out;
1231 	/*
1232 	 * Verify that no high bits are set.
1233 	 */
1234 	if (cpusetsize > sizeof(cpuset_t)) {
1235 		char *end;
1236 		char *cp;
1237 
1238 		end = cp = (char *)&mask->__bits;
1239 		end += cpusetsize;
1240 		cp += sizeof(cpuset_t);
1241 		while (cp != end)
1242 			if (*cp++ != 0) {
1243 				error = EINVAL;
1244 				goto out;
1245 			}
1246 
1247 	}
1248 	switch (level) {
1249 	case CPU_LEVEL_ROOT:
1250 	case CPU_LEVEL_CPUSET:
1251 		error = cpuset_which(which, id, &p, &ttd, &set);
1252 		if (error)
1253 			break;
1254 		switch (which) {
1255 		case CPU_WHICH_TID:
1256 		case CPU_WHICH_PID:
1257 			thread_lock(ttd);
1258 			set = cpuset_ref(ttd->td_cpuset);
1259 			thread_unlock(ttd);
1260 			PROC_UNLOCK(p);
1261 			break;
1262 		case CPU_WHICH_CPUSET:
1263 		case CPU_WHICH_JAIL:
1264 			break;
1265 		case CPU_WHICH_IRQ:
1266 		case CPU_WHICH_INTRHANDLER:
1267 		case CPU_WHICH_ITHREAD:
1268 		case CPU_WHICH_DOMAIN:
1269 			error = EINVAL;
1270 			goto out;
1271 		}
1272 		if (level == CPU_LEVEL_ROOT)
1273 			nset = cpuset_refroot(set);
1274 		else
1275 			nset = cpuset_refbase(set);
1276 		error = cpuset_modify(nset, mask);
1277 		cpuset_rel(nset);
1278 		cpuset_rel(set);
1279 		break;
1280 	case CPU_LEVEL_WHICH:
1281 		switch (which) {
1282 		case CPU_WHICH_TID:
1283 			error = cpuset_setthread(id, mask);
1284 			break;
1285 		case CPU_WHICH_PID:
1286 			error = cpuset_setproc(id, NULL, mask);
1287 			break;
1288 		case CPU_WHICH_CPUSET:
1289 		case CPU_WHICH_JAIL:
1290 			error = cpuset_which(which, id, &p, &ttd, &set);
1291 			if (error == 0) {
1292 				error = cpuset_modify(set, mask);
1293 				cpuset_rel(set);
1294 			}
1295 			break;
1296 		case CPU_WHICH_IRQ:
1297 		case CPU_WHICH_INTRHANDLER:
1298 		case CPU_WHICH_ITHREAD:
1299 			error = intr_setaffinity(id, which, mask);
1300 			break;
1301 		default:
1302 			error = EINVAL;
1303 			break;
1304 		}
1305 		break;
1306 	default:
1307 		error = EINVAL;
1308 		break;
1309 	}
1310 out:
1311 	free(mask, M_TEMP);
1312 	return (error);
1313 }
1314 
1315 #ifdef DDB
1316 void
1317 ddb_display_cpuset(const cpuset_t *set)
1318 {
1319 	int cpu, once;
1320 
1321 	for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) {
1322 		if (CPU_ISSET(cpu, set)) {
1323 			if (once == 0) {
1324 				db_printf("%d", cpu);
1325 				once = 1;
1326 			} else
1327 				db_printf(",%d", cpu);
1328 		}
1329 	}
1330 	if (once == 0)
1331 		db_printf("<none>");
1332 }
1333 
1334 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
1335 {
1336 	struct cpuset *set;
1337 
1338 	LIST_FOREACH(set, &cpuset_ids, cs_link) {
1339 		db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
1340 		    set, set->cs_id, set->cs_ref, set->cs_flags,
1341 		    (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
1342 		db_printf("  mask=");
1343 		ddb_display_cpuset(&set->cs_mask);
1344 		db_printf("\n");
1345 		if (db_pager_quit)
1346 			break;
1347 	}
1348 }
1349 #endif /* DDB */
1350